I tested 12 trading signals for grid bots on 6 years of data. Most of them were noise.

Alex Rincon — Thu, 05 Nov 2026 14:00:03 +0000

A grid bot buys and sells inside a price range. On futures you also choose a direction: long, short or neutral. Everyone who runs one asks the same thing every morning: which grid should I launch today?

I spent the last months trying to answer that with data instead of chart patterns. This is what I did, what failed, and what survived.

The setup

8 pairs (BTC, ETH, SOL, XRP, DOGE, BNB, ADA, LINK), Binance USDT perpetual futures.
4,846 simulated 30-day grid bots, one every 3 days per pair, from 2020 to September 2026. Each start is simulated three times: long, short and neutral.
1-minute candles. Before trusting the simulator I replayed 33 real bots and compared them order by order: 96.6% of the fills matched.

The question becomes measurable: for each start date, how much would choosing the right direction have paid (long − short), and can any signal known on that date predict it?

The rule I fixed before looking

This is the part that matters most. With enough candidate signals, something always "works" on past data. So the criterion was written down before running anything:

Spearman correlation with long − short, computed inside each pair.
Split history into three periods (2021–2023, 2024-05 → 2025-07, 2025-08 → now).
A signal passes only if it has the same sign and |ρ| ≥ 0.08 in all three.
Survivors are then tested walk-forward: every month, percentiles are recalibrated using only the past, exactly as it would run live.

for k, month in enumerate(months):
    past, cur = g[g.month < month], g[g.month == month]
    ref = np.sort(past[feature].dropna())  # past only
    pct = np.searchsorted(ref, cur[feature]) / len(ref)
    side = np.where(pct >= 2/3, "long",
           np.where(pct <= 1/3, "short", "neutral"))

What failed

Signal	2021–23	2024–25	2025–26
7-day momentum	+0.04	+0.01	−0.07	❌
30-day momentum	−0.01	−0.10	+0.05	❌
Fear & Greed index	+0.12	−0.10	−0.13	❌ flips sign
Change in top-trader positioning (1d / 7d)	~0	~0	~0	❌
RSI(6), daily	+0.01	+0.01	−0.04	❌
14-day breakout exhaustion ("reversal" labels)	+0.11	+0.11	−0.03	❌

The last two come from a popular "grid bot" indicator on TradingView. On the chart it looks great: it deletes its previous label every time it draws a new one, so you only ever see the latest call. Walk-forward, using it to pick the direction did −0.7 points per month worse than always launching neutral. I tried it on 1D, 4h and 1h candles: same result.

Fear & Greed is the interesting failure: it looked like a strong signal for two years and then reversed. That's exactly what the three-period rule is there to catch.

What survived

Only positioning data from Binance itself:

Top traders' long/short ratio — as a contrarian signal. When the biggest accounts are heavily long, the price tends to fall over the next weeks (ρ ≈ −0.10 to −0.25).
Aggressive buying (taker buy/sell volume) points the other way (ρ ≈ +0.12 to +0.20).

Combined (four measures of that positioning, each as a percentile of the pair's own history), walk-forward since 2022:

+1.8 points per 30-day bot versus always going neutral (95% CI +0.4 to +3.3).
Better in 32 of 52 months. Not 52. There are whole stretches where it does worse.

The range matters more than the direction

The surprise was the range. A fixed ±10% grid is what most people launch. Sizing it from the expected 30-day volatility instead:

	Fixed ±10%	Volatility-based
Bots that ended in profit	55%	70%
Worst 10% of bots	−24.1%	−11.5%

A wider grid earns less per day, but it loses far less when the market runs away — and the runaway months are what wipe out a grid's year.

One more rule fell out of the data: when a pair is in the calmest 20% of its past year, neutral grids lost money on average. Low volatility looks safe and isn't.

What I built with it

I turned the surviving rules into a small free page that recalculates every time you load it from live Binance data: direction, range and number of grids for BTC and ETH. The full validation — including the failures above — is published next to it:

Oracle: https://gridreplay.com/en/oracle/
Validation: https://gridreplay.com/en/validation/

Disclosure: I made it. The page has a Pionex affiliate link, and there's a paid tier planned for the other pairs. None of this is financial advice; it's an average edge over many months, not a prediction.

If you have a signal you think works for grids, tell me in the comments and I'll run it through the same test.